다중 선형회귀 분석 기반의 음성 존재 불확실성 추정 기법

Abstract

Conventional speech enhancement techniques use afixed ratio Q of the a priori probability of speechpresence and speech absence, or determine the valueof Q simply by comparing one particular parameteragainst a threshold in deriving the speech absenceprobability (SAP) associated with the speechpresence probability. To further improve theperformance of the SAP, we attempt to adaptivelychange Q according to a linear model consisting ofthe regression coefficients obtained by results frommultiple linear regression analysis and twoparameters: a priori SNR and the ratio between thelocal energy of the noisy speech and its derivedminimum.